Rethinking Depth Estimation for Multi-View Stereo: A Unified Representation and Focal Loss